# OMK MCP server

> An open-source multi-agent coding CLI that allows users to switch AI models mid-session and complete goals only when specified tests pass.

- Repository: https://github.com/dmae97/omk
- Page: https://hostedmcp.org/servers/dmae97/omk
- Transport: stdio
- Stars: 153
- License: MIT

## Install

### Claude Code

```sh
claude mcp add omk -- npx -y open-multi-agent-kit
```

### Cursor (`~/.cursor/mcp.json`)

```json
{
  "mcpServers": {
    "omk": {
      "command": "npx",
      "args": [
        "-y",
        "open-multi-agent-kit"
      ]
    }
  }
}
```

### VS Code (`.vscode/mcp.json`)

```json
{
  "servers": {
    "omk": {
      "type": "stdio",
      "command": "npx",
      "args": [
        "-y",
        "open-multi-agent-kit"
      ]
    }
  }
}
```

### VS Code CLI

```sh
code --add-mcp '{"name":"omk","type":"stdio","command":"npx","args":["-y","open-multi-agent-kit"]}'
```

### Claude Desktop (`claude_desktop_config.json`)

```json
{
  "mcpServers": {
    "omk": {
      "command": "npx",
      "args": [
        "-y",
        "open-multi-agent-kit"
      ]
    }
  }
}
```

### Codex

```sh
codex mcp add omk -- npx -y open-multi-agent-kit
```

### Codex config.toml (`~/.codex/config.toml`)

```toml
[mcp_servers.omk]
command = "npx"
args = ["-y", "open-multi-agent-kit"]
```

### Gemini CLI

```sh
gemini mcp add -s user omk npx -- -y open-multi-agent-kit
```

### Gemini settings.json (`~/.gemini/settings.json`)

```json
{
  "mcpServers": {
    "omk": {
      "command": "npx",
      "args": [
        "-y",
        "open-multi-agent-kit"
      ]
    }
  }
}
```

## Environment variables

| Name | Required | Secret | Description |
|---|---|---|---|
| `OMK_CONTEXT_GOVERNOR` | no | no |  |

## FAQ

### How do I ensure a goal is actually completed?

You can use the `/goal verify <command>` instruction to define a specific check that OMK must pass before the goal is marked as done.

### Can I change the AI model during a coding session?

Yes, you can switch providers and models within a single session using the `/model` command.

### Is the execution environment secure?

Yes, built-in bash commands run in a sandbox using sandbox-exec or bwrap, which lacks network access and fails closed.
